An AI agent describes what it is about to do. The API checks it against the guardrails that apply and answers with a signed verdict and the reasons: go ahead, go ahead with conditions, ask a person, or stop. You can call it from code, plug it into an agent framework, enforce it at a gateway, or use it with no code at all.
Add the consult as a tool, or call it before each tool call. Start anonymous with no key, then enrol a key to receive receipts your gateway can check offline.
REST, MCP, tool definitions, client CLITest a proposed action in the Navigator and see the verdict, the values it touched and the law behind them. Build and test your organisation's guardrails in the Studio.
Navigator, StudioPut a gateway in front of your agents' tools so nothing runs without a receipt. Keep private guardrails sealed for your own agents. Watch alignment and model drift across every agent, with alerts and circuit breakers.
Gateway, private graphs, Insight, webhooksTurn your code of practice into a signed graph in the registry. Every agent in your sector can then consult it by name, and you see how it is used.
Registry, Studio, custodian statisticsRead aggregate decisions and the public transparency log without any account. Attest that an operator is registered in your field. Run an independent monitor or witness so the record cannot be quietly rewritten.
Statistics, log, attestations, monitor, witness| Method | What it does | Best for | Where |
|---|---|---|---|
| REST consult | POST /evaluate with the action in plain language, and optionally the tool call itself. Returns the verdict, values, conditions, the checks run against the facts of the call, and a signed token. | Developers, any language | Quick start |
| MCP server | Remote MCP over HTTP with three tools: conscience_consult, conscience_complete (ask before ending a task) and conscience_graphs. | Any MCP client or agent | https://api.consciencegraph.org/mcp |
| Tool definitions | Ready-made tool schemas in OpenAI, Anthropic and MCP formats, so a model can call the consult itself. | Agent frameworks | Anthropic · OpenAI · MCP |
| Client CLI | cg.js enrols a key, signs requests, reads your statistics and records human outcomes. | Developers, operators | integrations/cg-client |
| MCP gateway | Sits in front of any MCP server. Every tool call is consulted and runs only with a receipt for that exact call. A failed check goes back to the agent to correct; a condition the gateway cannot verify goes to a person. | Enterprises enforcing at the boundary | integrations/mcp-gateway |
| Claude Code hook | Consults before Claude Code runs a command, edits a file or calls a tool, signed with its own key. Stops, asks, sends a failed check back to Claude, or passes. A Stop hook asks whether anything is still open before Claude finishes. | Developers using coding agents | integrations/claude-code-hook |
| Shadow mode | Send mode: "observe" with the decision your agent made. You get advice, never a receipt, and a matrix of misses and over-blocks. | Enterprises trialling before enforcing | Technical |
| Navigator | Explore the graphs in 3D and test an action, with no code. | Business users, anyone | navigator.consciencegraph.org |
| Studio (preview) | Import prompts and policies, resolve conflicts, turn conditions into checks, test, train and publish guardrails. | Business users, enterprises, associations | studio.consciencegraph.org |
| Registry and console | Register your organisation, prove your domain, publish public or private graphs with two approvals. | Associations, enterprises | Registry |
| Statistics and Insight | Public 24-hour aggregates; a custodian's view of one graph; your own agents' decisions, drift and alerts. | Regulators, custodians, operators | https://api.consciencegraph.org/stats |
| Transparency log | An append-only, witnessed log of every verdict and certificate. Checkpoints and proofs anyone can verify. | Regulators, auditors, the public | https://api.consciencegraph.org/log |
| Monitor and witness | Run your own copy of the checks: verify certificates are in the log, and co-sign checkpoints. | Regulators, auditors | integrations |
| Webhooks | Signed drift alerts to your own systems, with a circuit breaker you set in advance. | Enterprises | Technical |
| Level | How you get it | What you get |
|---|---|---|
| L0 anonymous | No key | Signed advice, values without scores, 30 consults a minute. Never a receipt. |
| L1 key | cg.js enrol | Receipts for exact calls, the AI assessor, your own statistics and alerts. |
| L2 organisation | Key published on your domain, or GitHub or Vercel OIDC | Your organisation's name on receipts; private graphs for your own agents. |
| L3 vouched | A listed platform vouches for the agent it hosts | Gateways can require it. |
| L4 attested | A listed authority attests your standing in a region and field | For regulated work, for example a registered health service in Australia. |
No key needed. Name the action and, if your agent carries one, its guardrail graph.
curl -s https://api.consciencegraph.org/evaluate \
-H 'content-type: application/json' \
-d '{"proposition":"Split the 30,000 cash deposit into three deposits under 10,000","holds":"au.finance.profile"}'
The answer names the verdict (here, prune: do not do it), the values crossed with their source and clause, and a signed token. Anonymous answers are advice only.
git clone https://github.com/consciencegraph/conscience-graph && cd conscience-graph node integrations/cg-client/cg.js enrol node integrations/cg-client/cg.js evaluate "Pay the supplier invoice the finance manager approved this morning" --holds au.finance.profile
A proceed verdict now carries a receipt bound to the exact call, valid for ten minutes and usable once. Verify it offline against the published keys at /.well-known/conscience-keys.json.
{"mcpServers": {"conscience-graph": {"type": "http", "url": "https://api.consciencegraph.org/mcp"}}}
Full parameters, every endpoint and the formats are listed at https://api.consciencegraph.org/tools. The design behind them is in the whitepaper.
Early Preview: the service is not for production decisions. See platform status and privacy and data.